static int get_addr(MPIDI_VC_t * vc, struct scif_portID *addr)
{
int mpi_errno = MPI_SUCCESS;
char *bc;
int pmi_errno;
int val_max_sz;
MPIU_CHKLMEM_DECL(1);
/* Allocate space for the business card */
pmi_errno = PMI_KVS_Get_value_length_max(&val_max_sz);
MPIU_ERR_CHKANDJUMP1(pmi_errno, mpi_errno, MPI_ERR_OTHER, "**fail",
"**fail %d", pmi_errno);
MPIU_CHKLMEM_MALLOC(bc, char *, val_max_sz, mpi_errno, "bc");
mpi_errno = vc->pg->getConnInfo(vc->pg_rank, bc, val_max_sz, vc->pg);
if (mpi_errno)
MPIU_ERR_POP(mpi_errno);
mpi_errno = scif_addr_from_bc(bc, &addr->node, &addr->port);
if (mpi_errno)
MPIU_ERR_POP(mpi_errno);
fn_exit:
MPIU_CHKLMEM_FREEALL();
return mpi_errno;
fn_fail:
goto fn_exit;
}
static int test_item3(void)
{
int rc = 0;
int val = 0;
val = random_value(10, 100);
if (PMI_SUCCESS != (rc = PMI_KVS_Get_key_length_max(&val))) {
log_fatal("PMI_KVS_Get_key_length_max failed: %d\n", rc);
return rc;
}
log_info("PMI_KVS_Get_key_length_max=%d\n", val);
if (!_legacy) {
log_assert(511 == val, "Check PMIX_MAX_KEYLEN value in pmix_common.h");
}
val = random_value(10, 100);
if (PMI_SUCCESS != (rc = PMI_KVS_Get_value_length_max(&val))) {
log_fatal("PMI_KVS_Get_value_length_max failed: %d\n", rc);
return rc;
}
log_info("PMI_KVS_Get_value_length_max=%d\n", val);
if (!_legacy) {
log_assert(4096 == val, "Check limitation for a value");
}
return rc;
}
int MPID_nem_ptl_init_id(MPIDI_VC_t *vc)
{
int mpi_errno = MPI_SUCCESS;
MPID_nem_ptl_vc_area *const vc_ptl = VC_PTL(vc);
char *bc;
int pmi_errno;
int val_max_sz;
MPIR_CHKLMEM_DECL(1);
MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_MPID_NEM_PTL_INIT_ID);
MPIR_FUNC_VERBOSE_ENTER(MPID_STATE_MPID_NEM_PTL_INIT_ID);
pmi_errno = PMI_KVS_Get_value_length_max(&val_max_sz);
MPIR_ERR_CHKANDJUMP1(pmi_errno, mpi_errno, MPI_ERR_OTHER, "**fail", "**fail %d", pmi_errno);
MPIR_CHKLMEM_MALLOC(bc, char *, val_max_sz, mpi_errno, "bc");
mpi_errno = vc->pg->getConnInfo(vc->pg_rank, bc, val_max_sz, vc->pg);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
mpi_errno = MPID_nem_ptl_get_id_from_bc(bc, &vc_ptl->id, &vc_ptl->pt, &vc_ptl->ptg, &vc_ptl->ptc, &vc_ptl->ptr, &vc_ptl->ptrg, &vc_ptl->ptrc);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
vc_ptl->id_initialized = TRUE;
MPIDI_CHANGE_VC_STATE(vc, ACTIVE);
fn_exit:
MPIR_CHKLMEM_FREEALL();
MPIR_FUNC_VERBOSE_EXIT(MPID_STATE_MPID_NEM_PTL_INIT_ID);
return mpi_errno;
fn_fail:
goto fn_exit;
}
static int pmi_set_proc_attr(const char* attr_name,
const void *buffer, size_t size)
{
char *attr, *attrval;
int rc;
if (NULL == pmi_kvs_name) {
int max_length;
rc = PMI_KVS_Get_value_length_max(&pmi_vallen_max);
if (PMI_SUCCESS != rc) {
ORTE_PMI_ERROR(rc, "PMI_Get_value_length_max");
return ORTE_ERROR;
}
if (PMI_SUCCESS != (rc = PMI_KVS_Get_name_length_max(&max_length))) {
ORTE_PMI_ERROR(rc, "PMI_KVS_Get_name_length_max");
return ORTE_ERROR;
}
pmi_kvs_name = malloc(max_length);
if (NULL == pmi_kvs_name) {
return ORTE_ERR_OUT_OF_RESOURCE;
}
rc = PMI_KVS_Get_my_name(pmi_kvs_name,max_length);
if (PMI_SUCCESS != rc) {
ORTE_PMI_ERROR(rc, "PMI_KVS_Get_my_name");
return ORTE_ERROR;
}
}
if (0 > asprintf(&attr, "%s-%s", ORTE_NAME_PRINT(ORTE_PROC_MY_NAME), attr_name)) {
return ORTE_ERR_OUT_OF_RESOURCE;
}
OPAL_OUTPUT_VERBOSE((1, orte_grpcomm_base.output,
"%s grpcomm:pmi: set attr %s of size %lu in KVS %s",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME), attr_name,
(unsigned long)size, pmi_kvs_name));
attrval = pmi_encode(buffer, size);
if (NULL == attrval) {
return ORTE_ERR_OUT_OF_RESOURCE;
}
if (strlen(attrval) > (size_t)pmi_vallen_max) {
opal_output(0, "pmi_proc_set_attr: attribute length is too long\n");
return ORTE_ERROR;
}
rc = PMI_KVS_Put(pmi_kvs_name, attr, attrval);
if (PMI_SUCCESS != rc) {
ORTE_PMI_ERROR(rc, "PMI_KVS_Put");
return ORTE_ERROR;
}
free(attr);
free(attrval);
return ORTE_SUCCESS;
}
int MPID_nem_mxm_vc_init(MPIDI_VC_t * vc)
{
int mpi_errno = MPI_SUCCESS;
MPIDI_CH3I_VC *vc_ch = &vc->ch;
MPID_nem_mxm_vc_area *vc_area = VC_BASE(vc);
MPIDI_STATE_DECL(MPID_STATE_MXM_VC_INIT);
MPIDI_FUNC_ENTER(MPID_STATE_MXM_VC_INIT);
/* local connection is used for any source communication */
MPIU_Assert(MPID_nem_mem_region.rank != vc->lpid);
MPIU_DBG_MSG_FMT(CH3_CHANNEL, VERBOSE,
(MPIU_DBG_FDEST,
"[%i]=== connecting to %i \n", MPID_nem_mem_region.rank, vc->lpid));
{
char *business_card;
int val_max_sz;
#ifdef USE_PMI2_API
val_max_sz = PMI2_MAX_VALLEN;
#else
mpi_errno = PMI_KVS_Get_value_length_max(&val_max_sz);
if (mpi_errno)
MPIU_ERR_POP(mpi_errno);
#endif
business_card = (char *) MPIU_Malloc(val_max_sz);
mpi_errno = vc->pg->getConnInfo(vc->pg_rank, business_card, val_max_sz, vc->pg);
if (mpi_errno)
MPIU_ERR_POP(mpi_errno);
vc_area->ctx = vc;
vc_area->mxm_ep = &_mxm_obj.endpoint[vc->pg_rank];
mpi_errno = _mxm_connect(&_mxm_obj.endpoint[vc->pg_rank], business_card, vc_area);
if (mpi_errno)
MPIU_ERR_POP(mpi_errno);
MPIU_Free(business_card);
}
MPIDI_CHANGE_VC_STATE(vc, ACTIVE);
vc_area->pending_sends = 0;
vc->rndvSend_fn = NULL;
vc->rndvRecv_fn = NULL;
vc->sendNoncontig_fn = MPID_nem_mxm_SendNoncontig;
vc->comm_ops = &comm_ops;
vc_ch->iStartContigMsg = MPID_nem_mxm_iStartContigMsg;
vc_ch->iSendContig = MPID_nem_mxm_iSendContig;
fn_exit:
MPIDI_FUNC_EXIT(MPID_STATE_MXM_VC_INIT);
return mpi_errno;
fn_fail:
goto fn_exit;
}
int
MPID_nem_mx_vc_init (MPIDI_VC_t *vc)
{
uint32_t threshold;
MPIDI_CH3I_VC *vc_ch = VC_CH(vc);
int mpi_errno = MPI_SUCCESS;
/* first make sure that our private fields in the vc fit into the area provided */
MPIU_Assert(sizeof(MPID_nem_mx_vc_area) <= MPID_NEM_VC_NETMOD_AREA_LEN);
#ifdef ONDEMAND
VC_FIELD(vc, local_connected) = 0;
VC_FIELD(vc, remote_connected) = 0;
#else
{
char *business_card;
int val_max_sz;
int ret;
#ifdef USE_PMI2_API
val_max_sz = PMI2_MAX_VALLEN;
#else
mpi_errno = PMI_KVS_Get_value_length_max(&val_max_sz);
#endif
business_card = (char *)MPIU_Malloc(val_max_sz);
mpi_errno = vc->pg->getConnInfo(vc->pg_rank, business_card,val_max_sz, vc->pg);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
mpi_errno = MPID_nem_mx_get_from_bc (business_card, &VC_FIELD(vc, remote_endpoint_id), &VC_FIELD(vc, remote_nic_id));
if (mpi_errno) MPIU_ERR_POP (mpi_errno);
MPIU_Free(business_card);
ret = mx_connect(MPID_nem_mx_local_endpoint,VC_FIELD(vc, remote_nic_id),VC_FIELD(vc, remote_endpoint_id),
MPID_NEM_MX_FILTER,MX_INFINITE,&(VC_FIELD(vc, remote_endpoint_addr)));
MPIU_ERR_CHKANDJUMP1 (ret != MX_SUCCESS, mpi_errno, MPI_ERR_OTHER, "**mx_connect", "**mx_connect %s", mx_strerror (ret));
mx_set_endpoint_addr_context(VC_FIELD(vc, remote_endpoint_addr),(void *)vc);
MPIDI_CHANGE_VC_STATE(vc, ACTIVE);
}
#endif
mx_get_info(MPID_nem_mx_local_endpoint, MX_COPY_SEND_MAX, NULL, 0, &threshold, sizeof(uint32_t));
vc->eager_max_msg_sz = threshold;
vc->rndvSend_fn = NULL;
vc->sendNoncontig_fn = MPID_nem_mx_SendNoncontig;
vc->comm_ops = &comm_ops;
vc_ch->iStartContigMsg = MPID_nem_mx_iStartContigMsg;
vc_ch->iSendContig = MPID_nem_mx_iSendContig;
fn_exit:
return mpi_errno;
fn_fail:
goto fn_exit;
}
static int setup_pmi(void)
{
int max_length, rc;
#if WANT_CRAY_PMI2_EXT
pmi_vallen_max = PMI2_MAX_VALLEN;
#else
rc = PMI_KVS_Get_value_length_max(&pmi_vallen_max);
if (PMI_SUCCESS != rc) {
ORTE_PMI_ERROR(rc, "PMI_Get_value_length_max");
return ORTE_ERROR;
}
#endif
pmi_attr_val = malloc(pmi_vallen_max);
if (NULL == pmi_attr_val) {
return ORTE_ERR_OUT_OF_RESOURCE;
}
#if WANT_CRAY_PMI2_EXT
/* TODO -- is this ok */
max_length = 1024;
#else
if (PMI_SUCCESS != (rc = PMI_KVS_Get_name_length_max(&max_length))) {
ORTE_PMI_ERROR(rc, "PMI_KVS_Get_name_length_max");
return ORTE_ERROR;
}
#endif
pmi_kvs_name = malloc(max_length);
if (NULL == pmi_kvs_name) {
return ORTE_ERR_OUT_OF_RESOURCE;
}
#if WANT_CRAY_PMI2_EXT
rc = PMI2_Job_GetId(pmi_kvs_name, max_length);
#else
rc = PMI_KVS_Get_my_name(pmi_kvs_name,max_length);
#endif
if (PMI_SUCCESS != rc) {
ORTE_PMI_ERROR(rc, "PMI_KVS_Get_my_name");
return ORTE_ERROR;
}
#if WANT_CRAY_PMI2_EXT
pmi_keylen_max = PMI2_MAX_KEYLEN;
#else
if (PMI_SUCCESS != (rc = PMI_KVS_Get_key_length_max(&pmi_keylen_max))) {
ORTE_PMI_ERROR(rc, "PMI_KVS_Get_key_length_max");
return ORTE_ERROR;
}
#endif
pmi_kvs_key = malloc(pmi_keylen_max);
return ORTE_SUCCESS;
}
int MPIDI_check_for_failed_procs(void)
{
int mpi_errno = MPI_SUCCESS;
int pmi_errno;
int len;
char *kvsname = MPIDI_global.jobid;
char *failed_procs_string = NULL;
MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_MPIDI_CHECK_FOR_FAILED_PROCS);
MPIR_FUNC_VERBOSE_ENTER(MPID_STATE_MPIDI_CHECK_FOR_FAILED_PROCS);
/* FIXME: Currently this only handles failed processes in
* comm_world. We need to fix hydra to include the pgid along
* with the rank, then we need to create the failed group from
* something bigger than comm_world. */
#ifdef USE_PMIX_API
MPIR_Assert(0);
#elif defined(USE_PMI2_API)
{
int vallen = 0;
len = PMI2_MAX_VALLEN;
failed_procs_string = MPL_malloc(len, MPL_MEM_OTHER);
MPIR_Assert(failed_procs_string);
pmi_errno =
PMI2_KVS_Get(kvsname, PMI2_ID_NULL, "PMI_dead_processes", failed_procs_string,
len, &vallen);
MPIR_ERR_CHKANDJUMP(pmi_errno, mpi_errno, MPI_ERR_OTHER, "**pmi_kvs_get");
MPL_free(failed_procs_string);
}
#else
pmi_errno = PMI_KVS_Get_value_length_max(&len);
MPIR_ERR_CHKANDJUMP(pmi_errno, mpi_errno, MPI_ERR_OTHER, "**pmi_kvs_get_value_length_max");
failed_procs_string = MPL_malloc(len, MPL_MEM_OTHER);
MPIR_Assert(failed_procs_string);
pmi_errno = PMI_KVS_Get(kvsname, "PMI_dead_processes", failed_procs_string, len);
MPIR_ERR_CHKANDJUMP(pmi_errno, mpi_errno, MPI_ERR_OTHER, "**pmi_kvs_get");
MPL_free(failed_procs_string);
#endif
MPL_DBG_MSG_FMT(MPIDI_CH4_DBG_GENERAL, VERBOSE,
(MPL_DBG_FDEST, "Received proc fail notification: %s", failed_procs_string));
/* FIXME: handle ULFM failed groups here */
fn_exit:
MPIR_FUNC_VERBOSE_EXIT(MPID_STATE_MPIDI_CHECK_FOR_FAILED_PROCS);
return mpi_errno;
fn_fail:
MPL_free(failed_procs_string);
goto fn_exit;
}
int
MPID_nem_newmad_vc_init (MPIDI_VC_t *vc)
{
MPIDI_CH3I_VC *vc_ch = VC_CH(vc);
char *business_card;
int mpi_errno = MPI_SUCCESS;
int val_max_sz;
int ret;
#ifdef USE_PMI2_API
val_max_sz = PMI2_MAX_VALLEN;
#else
mpi_errno = PMI_KVS_Get_value_length_max(&val_max_sz);
#endif
business_card = (char *)MPIU_Malloc(val_max_sz);
mpi_errno = vc->pg->getConnInfo(vc->pg_rank, business_card,val_max_sz,vc->pg);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
/* Very important */
memset(VC_FIELD(vc, url),0,MPID_NEM_NMAD_MAX_SIZE);
mpi_errno = MPID_nem_newmad_get_from_bc (business_card, VC_FIELD(vc, url));
if (mpi_errno) MPIU_ERR_POP (mpi_errno);
MPIU_Free(business_card);
ret = nm_session_connect(mpid_nem_newmad_session, &(VC_FIELD(vc,p_gate)), VC_FIELD(vc, url));
if (ret != NM_ESUCCESS) fprintf(stdout,"nm_session_connect returned ret = %d\n", ret);
nm_gate_ref_set(VC_FIELD(vc, p_gate),(void*)vc);
MPIDI_CHANGE_VC_STATE(vc, ACTIVE);
vc->eager_max_msg_sz = 32768;
vc->rndvSend_fn = NULL;
vc->sendNoncontig_fn = MPID_nem_newmad_SendNoncontig;
vc->comm_ops = &comm_ops;
vc_ch->iStartContigMsg = MPID_nem_newmad_iStartContigMsg;
vc_ch->iSendContig = MPID_nem_newmad_iSendContig;
fn_exit:
return mpi_errno;
fn_fail:
goto fn_exit;
}
static int setup_pmi(void)
{
int max_length, rc;
#if WANT_PMI2_SUPPORT
pmi_vallen_max = PMI2_MAX_VALLEN;
#else
rc = PMI_KVS_Get_value_length_max(&pmi_vallen_max);
if (PMI_SUCCESS != rc) {
return OMPI_ERROR;
}
#endif
#if WANT_PMI2_SUPPORT
/* TODO -- is this ok */
max_length = 1024;
#else
if (PMI_SUCCESS != (rc = PMI_KVS_Get_name_length_max(&max_length))) {
return OMPI_ERROR;
}
#endif
pmi_kvs_name = (char*)malloc(max_length);
if (NULL == pmi_kvs_name) {
return OMPI_ERR_OUT_OF_RESOURCE;
}
#if WANT_PMI2_SUPPORT
rc = PMI2_Job_GetId(pmi_kvs_name, max_length);
#else
rc = PMI_KVS_Get_my_name(pmi_kvs_name,max_length);
#endif
if (PMI_SUCCESS != rc) {
return OMPI_ERROR;
}
#if WANT_PMI2_SUPPORT
pmi_keylen_max = PMI2_MAX_KEYLEN;
#else
if (PMI_SUCCESS != (rc = PMI_KVS_Get_key_length_max(&pmi_keylen_max))) {
return OMPI_ERROR;
}
#endif
return OMPI_SUCCESS;
}
static int test_item5(void)
{
int rc = 0;
char *val = NULL;
int val_size = 0;
/* Predefined Job attributes */
const char *tkeys[] = {
"PMI_process_mapping",
NULL
};
const char **ptr = tkeys;
if (_legacy || !_legacy) {
log_error("%s\n", "PMIx and SLURM/PMI1 do not set 'PMI_process_mapping' (Do not mark test as failed)");
return rc;
}
if (PMI_SUCCESS != (rc = PMI_KVS_Get_value_length_max(&val_size))) {
log_fatal("PMI_KVS_Get_value_length_max failed: %d\n", rc);
return rc;
}
val = alloca(val_size);
if (!val) {
return PMI_FAIL;
}
while (*ptr) {
if (PMI_SUCCESS != (rc = PMI_KVS_Get(jobid, *ptr, val, val_size))) {
log_fatal("PMI_KVS_Get: [%s] %d\n", *ptr, rc);
return rc;
}
log_info("key=%s value=%.80s\n", *ptr, val);
ptr++;
}
return rc;
}
/* gasneti_bootstrapInit
*/
int gasneti_bootstrapInit_pmi(
int *argc_p, char ***argv_p,
gasnet_node_t *nodes_p, gasnet_node_t *mynode_p) {
int size, rank;
#if USE_PMI2_API
int spawned, appnum;
if (PMI2_SUCCESS != PMI2_Init(&spawned, &size, &rank, &appnum)) {
return GASNET_ERR_NOT_INIT;
}
#else
if (PMI_SUCCESS != PMI_Initialized(&gasneti_pmi_initialized)) {
return GASNET_ERR_NOT_INIT;
}
if (PMI_FALSE == gasneti_pmi_initialized) {
int spawned;
if (PMI_SUCCESS != PMI_Init(&spawned)) {
return GASNET_ERR_NOT_INIT;
}
}
if (PMI_SUCCESS != PMI_Get_rank(&rank)) {
gasneti_fatalerror("PMI_Get_rank() failed");
}
if (PMI_SUCCESS != PMI_Get_size(&size)) {
gasneti_fatalerror("PMI_Get_size() failed");
}
#endif
*mynode_p = rank;
*nodes_p = size;
#if USE_PMI2_API
max_name_len = 1024; /* XXX: can almost certainly be shorter than this! */
max_key_len = PMI2_MAX_KEYLEN;
max_val_len = PMI2_MAX_VALLEN;
#else
if (PMI_SUCCESS != PMI_KVS_Get_name_length_max(&max_name_len)) {
gasneti_fatalerror("PMI_KVS_Get_name_length_max() failed");
}
if (PMI_SUCCESS != PMI_KVS_Get_key_length_max(&max_key_len)) {
gasneti_fatalerror("PMI_KVS_Get_key_length_max() failed");
}
if (PMI_SUCCESS != PMI_KVS_Get_value_length_max(&max_val_len)) {
gasneti_fatalerror("PMI_KVS_Get_value_length_max() failed");
}
#endif
kvs_name = (char*) gasneti_malloc(max_name_len);
kvs_key = (char*) gasneti_malloc(max_key_len);
kvs_value = (char*) gasneti_malloc(max_val_len);
max_val_bytes = 4 * (max_val_len / 5);
#if USE_PMI2_API
if (PMI2_SUCCESS != PMI2_Job_GetId(kvs_name, max_name_len)) {
gasneti_fatalerror("PMI2_Job_GetId() failed");
}
#else
if (PMI_SUCCESS != PMI_KVS_Get_my_name(kvs_name, max_name_len)) {
gasneti_fatalerror("PMI_KVS_Get_my_name() failed");
}
#endif
return GASNET_OK;
}
int
main(int argc,
char *argv[])
{
int initialized, rank, size;
int i, max_name_len, max_key_len, max_val_len;
char *name, *key, *val;
if (PMI_SUCCESS != PMI_Initialized(&initialized)) {
return 1;
}
if (0 == initialized) {
if (PMI_SUCCESS != PMI_Init(&initialized)) {
return 1;
}
}
if (PMI_SUCCESS != PMI_Get_rank(&rank)) {
return 1;
}
if (PMI_SUCCESS != PMI_Get_size(&size)) {
return 1;
}
printf("Hello, World. I am %d of %d\n", rank, size);
if (PMI_SUCCESS != PMI_KVS_Get_name_length_max(&max_name_len)) {
return 1;
}
name = (char*) malloc(max_name_len);
if (NULL == name) return 1;
if (PMI_SUCCESS != PMI_KVS_Get_key_length_max(&max_key_len)) {
return 1;
}
key = (char*) malloc(max_key_len);
if (NULL == key) return 1;
if (PMI_SUCCESS != PMI_KVS_Get_value_length_max(&max_val_len)) {
return 1;
}
val = (char*) malloc(max_val_len);
if (NULL == val) return 1;
if (PMI_SUCCESS != PMI_KVS_Get_my_name(name, max_name_len)) {
return 1;
}
/* put my information */
snprintf(key, max_key_len, "pmi_hello-%lu-test", (long unsigned) rank);
snprintf(val, max_val_len, "%lu", (long unsigned) rank);
if (PMI_SUCCESS != PMI_KVS_Put(name, key, val)) {
return 1;
}
if (PMI_SUCCESS != PMI_KVS_Commit(name)) {
return 1;
}
if (PMI_SUCCESS != PMI_Barrier()) {
return 1;
}
/* verify everyone's information */
for (i = 0 ; i < size ; ++i) {
snprintf(key, max_key_len, "pmi_hello-%lu-test", (long unsigned) i);
if (PMI_SUCCESS != PMI_KVS_Get(name, key, val, max_val_len)) {
return 1;
}
if (i != strtol(val, NULL, 0)) {
fprintf(stderr, "%d: Error: Expected %d, got %d\n", rank, i, (int) strtol(val, NULL, 0));
return 1;
}
}
PMI_Finalize();
return 0;
}
static int s1_init(void)
{
PMI_BOOL initialized;
int spawned;
int rc, ret = OPAL_ERROR;
int i;
char *pmix_id, *tmp;
uint32_t jobfam, stepid;
opal_value_t kv;
if (PMI_SUCCESS != (rc = PMI_Initialized(&initialized))) {
OPAL_PMI_ERROR(rc, "PMI_Initialized");
return OPAL_ERROR;
}
if( PMI_TRUE != initialized && PMI_SUCCESS != (rc = PMI_Init(&spawned)) ) {
OPAL_PMI_ERROR(rc, "PMI_Init");
return OPAL_ERROR;
}
// Initialize space demands
rc = PMI_KVS_Get_value_length_max(&pmix_vallen_max);
if( PMI_SUCCESS != rc ) {
OPAL_PMI_ERROR(rc, "PMI_KVS_Get_value_length_max");
goto err_exit;
}
rc = PMI_KVS_Get_name_length_max(&pmix_kvslen_max);
if (PMI_SUCCESS != rc ) {
OPAL_PMI_ERROR(rc, "PMI_KVS_Get_name_length_max");
goto err_exit;
}
rc = PMI_KVS_Get_key_length_max(&pmix_keylen_max);
if( PMI_SUCCESS != rc ) {
OPAL_PMI_ERROR(rc, "PMI_KVS_Get_key_length_max");
goto err_exit;
}
// Initialize job environment information
pmix_id = (char*)malloc(pmix_vallen_max);
if( pmix_id == NULL ){
ret = OPAL_ERR_OUT_OF_RESOURCE;
goto err_exit;
}
/* Get domain id */
if (PMI_SUCCESS != (rc = PMI_Get_kvs_domain_id(pmix_id, pmix_vallen_max))) {
free(pmix_id);
goto err_exit;
}
/* Slurm PMI provides the job id as an integer followed
* by a '.', followed by essentially a stepid. The first integer
* defines an overall job number. The second integer is the number of
* individual jobs we have run within that allocation. So we translate
* this as the overall job number equating to our job family, and
* the individual number equating to our local jobid
*/
jobfam = strtoul(pmix_id, &tmp, 10);
if (NULL == tmp) {
/* hmmm - no '.', so let's just use zero */
stepid = 0;
} else {
tmp++; /* step over the '.' */
stepid = strtoul(tmp, NULL, 10);
}
/* now build the jobid */
s1_jobid = (jobfam << 16) | stepid;
free(pmix_id);
/* get our rank */
ret = PMI_Get_rank(&s1_rank);
if( PMI_SUCCESS != ret ) {
OPAL_PMI_ERROR(ret, "PMI_Get_rank");
goto err_exit;
}
/* store our name in the opal_proc_t so that
* debug messages will make sense - an upper
* layer will eventually overwrite it, but that
* won't do any harm */
s1_pname.jid = s1_jobid;
s1_pname.vid = s1_rank;
opal_proc_set_name((opal_process_name_t*)&s1_pname);
opal_output_verbose(2, opal_pmix_base_framework.framework_output,
"%s pmix:s1: assigned tmp name",
OPAL_NAME_PRINT(*(opal_process_name_t*)&s1_pname));
pmix_kvs_name = (char*)malloc(pmix_kvslen_max);
if( pmix_kvs_name == NULL ){
ret = OPAL_ERR_OUT_OF_RESOURCE;
goto err_exit;
}
rc = PMI_KVS_Get_my_name(pmix_kvs_name, pmix_kvslen_max);
if( PMI_SUCCESS != rc ) {
OPAL_PMI_ERROR(rc, "PMI_KVS_Get_my_name");
goto err_exit;
}
/* get our local proc info to find our local rank */
if (PMI_SUCCESS != (rc = PMI_Get_clique_size(&s1_nlranks))) {
OPAL_PMI_ERROR(rc, "PMI_Get_clique_size");
return rc;
}
/* now get the specific ranks */
s1_lranks = (int*)calloc(s1_nlranks, sizeof(int));
if (NULL == s1_lranks) {
rc = OPAL_ERR_OUT_OF_RESOURCE;
OPAL_ERROR_LOG(rc);
return rc;
}
if (PMI_SUCCESS != (rc = PMI_Get_clique_ranks(s1_lranks, s1_nlranks))) {
OPAL_PMI_ERROR(rc, "PMI_Get_clique_ranks");
free(s1_lranks);
return rc;
}
/* find ourselves */
for (i=0; i < s1_nlranks; i++) {
if (s1_rank == s1_lranks[i]) {
s1_lrank = i;
s1_nrank = i;
break;
}
}
/* get universe size */
ret = PMI_Get_universe_size(&s1_usize);
if (PMI_SUCCESS != ret) {
OPAL_PMI_ERROR(ret, "PMI_Get_universe_size");
goto err_exit;
}
/* push this into the dstore for subsequent fetches */
OBJ_CONSTRUCT(&kv, opal_value_t);
kv.key = strdup(OPAL_DSTORE_UNIV_SIZE);
kv.type = OPAL_UINT32;
kv.data.uint32 = s1_usize;
if (OPAL_SUCCESS != (ret = opal_dstore.store(opal_dstore_internal, &OPAL_PROC_MY_NAME, &kv))) {
OPAL_ERROR_LOG(ret);
OBJ_DESTRUCT(&kv);
goto err_exit;
}
OBJ_DESTRUCT(&kv);
/* get job size */
ret = PMI_Get_size(&s1_jsize);
if (PMI_SUCCESS != ret) {
OPAL_PMI_ERROR(ret, "PMI_Get_size");
goto err_exit;
}
/* get appnum */
ret = PMI_Get_appnum(&s1_appnum);
if (PMI_SUCCESS != ret) {
OPAL_PMI_ERROR(ret, "PMI_Get_appnum");
goto err_exit;
}
return OPAL_SUCCESS;
err_exit:
PMI_Finalize();
return ret;
}
int MPIDU_bc_table_create(int rank, int size, int *nodemap, void *bc, int bc_len, int same_len,
int roots_only, void **bc_table, size_t ** bc_indices)
{
int rc, mpi_errno = MPI_SUCCESS;
int start, end, i;
int key_max, val_max, name_max, out_len, rem;
char *kvsname = NULL, *key = NULL, *val = NULL, *val_p;
int local_rank = -1, local_leader = -1;
size_t my_bc_len = bc_len;
MPIR_NODEMAP_get_local_info(rank, size, nodemap, &local_size, &local_rank, &local_leader);
rc = PMI_KVS_Get_name_length_max(&name_max);
MPIR_ERR_CHKANDJUMP(rc, mpi_errno, MPI_ERR_OTHER, "**pmi_kvs_get_name_length_max");
rc = PMI_KVS_Get_key_length_max(&key_max);
MPIR_ERR_CHKANDJUMP(rc, mpi_errno, MPI_ERR_OTHER, "**pmi_kvs_get_key_length_max");
rc = PMI_KVS_Get_value_length_max(&val_max);
MPIR_ERR_CHKANDJUMP(rc, mpi_errno, MPI_ERR_OTHER, "**pmi_kvs_get_value_length_max");
/* if business cards can be different length, use the max value length */
if (!same_len)
bc_len = val_max;
mpi_errno = MPIDU_shm_seg_alloc(bc_len * size, (void **) &segment, MPL_MEM_ADDRESS);
if (mpi_errno)
MPIR_ERR_POP(mpi_errno);
mpi_errno =
MPIDU_shm_seg_commit(&memory, &barrier, local_size, local_rank, local_leader, rank,
MPL_MEM_ADDRESS);
if (mpi_errno)
MPIR_ERR_POP(mpi_errno);
if (size == 1) {
memcpy(segment, bc, my_bc_len);
goto single;
}
kvsname = MPL_malloc(name_max, MPL_MEM_ADDRESS);
MPIR_Assert(kvsname);
rc = PMI_KVS_Get_my_name(kvsname, name_max);
MPIR_ERR_CHKANDJUMP(rc, mpi_errno, MPI_ERR_OTHER, "**pmi_kvs_get_my_name");
val = MPL_malloc(val_max, MPL_MEM_ADDRESS);
memset(val, 0, val_max);
val_p = val;
rem = val_max;
rc = MPL_str_add_binary_arg(&val_p, &rem, "mpi", (char *) bc, my_bc_len);
MPIR_ERR_CHKANDJUMP(rc, mpi_errno, MPI_ERR_OTHER, "**buscard");
MPIR_Assert(rem >= 0);
key = MPL_malloc(key_max, MPL_MEM_ADDRESS);
MPIR_Assert(key);
if (!roots_only || rank == local_leader) {
sprintf(key, "bc-%d", rank);
rc = PMI_KVS_Put(kvsname, key, val);
MPIR_ERR_CHKANDJUMP(rc, mpi_errno, MPI_ERR_OTHER, "**pmi_kvs_put");
rc = PMI_KVS_Commit(kvsname);
MPIR_ERR_CHKANDJUMP(rc, mpi_errno, MPI_ERR_OTHER, "**pmi_kvs_commit");
}
rc = PMI_Barrier();
MPIR_ERR_CHKANDJUMP(rc, mpi_errno, MPI_ERR_OTHER, "**pmi_barrier");
if (!roots_only) {
start = local_rank * (size / local_size);
end = start + (size / local_size);
if (local_rank == local_size - 1)
end += size % local_size;
for (i = start; i < end; i++) {
sprintf(key, "bc-%d", i);
rc = PMI_KVS_Get(kvsname, key, val, val_max);
MPIR_ERR_CHKANDJUMP(rc, mpi_errno, MPI_ERR_OTHER, "**pmi_kvs_get");
rc = MPL_str_get_binary_arg(val, "mpi", &segment[i * bc_len], bc_len, &out_len);
MPIR_ERR_CHKANDJUMP(rc, mpi_errno, MPI_ERR_OTHER, "**argstr_missinghost");
}
} else {
int num_nodes, *node_roots;
MPIR_NODEMAP_get_node_roots(nodemap, size, &node_roots, &num_nodes);
start = local_rank * (num_nodes / local_size);
end = start + (num_nodes / local_size);
if (local_rank == local_size - 1)
end += num_nodes % local_size;
for (i = start; i < end; i++) {
sprintf(key, "bc-%d", node_roots[i]);
rc = PMI_KVS_Get(kvsname, key, val, val_max);
MPIR_ERR_CHKANDJUMP(rc, mpi_errno, MPI_ERR_OTHER, "**pmi_kvs_get");
rc = MPL_str_get_binary_arg(val, "mpi", &segment[i * bc_len], bc_len, &out_len);
MPIR_ERR_CHKANDJUMP(rc, mpi_errno, MPI_ERR_OTHER, "**argstr_missinghost");
}
MPL_free(node_roots);
}
mpi_errno = MPIDU_shm_barrier(barrier, local_size);
if (mpi_errno)
MPIR_ERR_POP(mpi_errno);
single:
if (!same_len) {
indices = MPL_malloc(size * sizeof(size_t), MPL_MEM_ADDRESS);
MPIR_Assert(indices);
for (i = 0; i < size; i++)
indices[i] = bc_len * i;
*bc_indices = indices;
}
fn_exit:
MPL_free(kvsname);
MPL_free(key);
MPL_free(val);
*bc_table = segment;
return mpi_errno;
fn_fail:
goto fn_exit;
}
int main(int argc, char **argv, char **envp)
{
int pmi_rank = -1;
int pmi_process_group_size = -1;
int num_local_procs = 0;
int *local_rank_ids = NULL;
int spawned = PMI_FALSE;
int rc = EXIT_SUCCESS;
char *err = NULL;
PMI_BOOL pmi_initialized = PMI_FALSE;
int pmi_vallen_max, max_length;
char *pmi_kvs_name;
/* sanity */
if (PMI_SUCCESS != PMI_Initialized(&pmi_initialized) ||
PMI_TRUE == pmi_initialized) {
fprintf(stderr, "=== ERROR: PMI sanity failure\n");
return EXIT_FAILURE;
}
if (PMI_SUCCESS != PMI_Init(&spawned)) {
err = "PMI_Init failure!";
goto done;
}
if (PMI_SUCCESS != PMI_Get_size(&pmi_process_group_size)) {
err = "PMI_Get_size failure!";
goto done;
}
if (PMI_SUCCESS != PMI_Get_rank(&pmi_rank)) {
err = "PMI_Get_rank failure!";
goto done;
}
if (PMI_SUCCESS != PMI_Get_clique_size(&num_local_procs)) {
err = "PMI_Get_clique_size failure!";
goto done;
}
if (PMI_SUCCESS != PMI_KVS_Get_value_length_max(&pmi_vallen_max)) {
err = "PMI_KVS_Get_value_length_max failure!";
goto done;
}
if (PMI_SUCCESS != PMI_KVS_Get_name_length_max(&max_length)) {
err = "PMI_KVS_Get_name_length_max failure!";
goto done;
}
pmi_kvs_name = (char*)malloc(max_length);
if (NULL == pmi_kvs_name) {
err = "malloc failure!";
goto done;
}
if (PMI_SUCCESS != PMI_KVS_Get_my_name(pmi_kvs_name,max_length)) {
err = "PMI_KVS_Get_my_name failure!";
goto done;
}
if (NULL == (local_rank_ids = calloc(num_local_procs, sizeof(int)))) {
err = "out of resources";
goto done;
}
if (PMI_SUCCESS != PMI_Get_clique_ranks(local_rank_ids, num_local_procs)) {
err = "PMI_Get_clique_size failure!";
goto done;
}
/* lowest local rank will print env info and tag its output*/
// if (pmi_rank == local_rank_ids[0]) {
// for (; NULL != envp && NULL != *envp; ++envp) {
// printf("===[%d]: %s\n", pmi_rank, *envp);
// }
//}
done:
if (PMI_TRUE == pmi_initialized) {
if (PMI_SUCCESS != PMI_Finalize()) {
err = "PMI_Finalize failure!";
}
}
if (NULL != err) {
fprintf(stderr, "=== ERROR [rank:%d] %s\n", pmi_rank, err);
rc = EXIT_FAILURE;
}
return rc;
}
main (int argc, char **argv)
{
int i, j, rc;
int nprocs, procid;
int clique_size, *clique_ranks = NULL;
char *jobid_ptr, *nprocs_ptr, *procid_ptr;
int pmi_rank, pmi_size, kvs_name_len, key_len, val_len;
PMI_BOOL initialized;
char *key, *val, *kvs_name;
struct timeval tv1, tv2;
long delta_t;
char tv_str[20];
gettimeofday(&tv1, NULL);
/* Get process count and our id from environment variables */
jobid_ptr = getenv("SLURM_JOB_ID");
nprocs_ptr = getenv("SLURM_NPROCS");
procid_ptr = getenv("SLURM_PROCID");
if (jobid_ptr == NULL) {
printf("WARNING: PMI test not run under SLURM\n");
nprocs = 1;
procid = 0;
} else if ((nprocs_ptr == NULL) || (procid_ptr == NULL)) {
printf("FAILURE: SLURM environment variables not set\n");
exit(1);
} else {
nprocs = atoi(nprocs_ptr);
procid = atoi(procid_ptr);
}
/* Validate process count and our id */
if ((nprocs < 1) || (nprocs > 9999)) {
printf("FAILURE: Invalid nprocs %s\n", nprocs_ptr);
exit(1);
}
if ((procid < 0) || (procid > 9999)) {
printf("FAILURE: Invalid procid %s\n", procid_ptr);
exit(1);
}
/* Get process count and size from PMI and validate */
if ((rc = PMI_Init(&i)) != PMI_SUCCESS) {
printf("FAILURE: PMI_Init: %d\n", rc);
exit(1);
}
initialized = PMI_FALSE;
if ((rc = PMI_Initialized(&initialized)) != PMI_SUCCESS) {
printf("FAILURE: PMI_Initialized: %d\n", rc);
exit(1);
}
if (initialized != PMI_TRUE) {
printf("FAILURE: PMI_Initialized returned false\n");
exit(1);
}
if ((rc = PMI_Get_rank(&pmi_rank)) != PMI_SUCCESS) {
printf("FAILURE: PMI_Get_rank: %d\n", rc);
exit(1);
}
#if _DEBUG
printf("PMI_Get_rank = %d\n", pmi_rank);
#endif
if ((rc = PMI_Get_size(&pmi_size)) != PMI_SUCCESS) {
printf("FAILURE: PMI_Get_size: %d, task %d\n", rc, pmi_rank);
exit(1);
}
#if _DEBUG
printf("PMI_Get_size = %d\n", pmi_size);
#endif
if (pmi_rank != procid) {
printf("FAILURE: Rank(%d) != PROCID(%d)\n",
pmi_rank, procid);
exit(1);
}
if (pmi_size != nprocs) {
printf("FAILURE: Size(%d) != NPROCS(%d), task %d\n",
pmi_size, nprocs, pmi_rank);
exit(1);
}
if ((rc = PMI_Get_clique_size(&clique_size)) != PMI_SUCCESS) {
printf("FAILURE: PMI_Get_clique_size: %d, task %d\n",
rc, pmi_rank);
exit(1);
}
clique_ranks = malloc(sizeof(int) * clique_size);
if ((rc = PMI_Get_clique_ranks(clique_ranks, clique_size)) !=
PMI_SUCCESS) {
printf("FAILURE: PMI_Get_clique_ranks: %d, task %d\n",
rc, pmi_rank);
exit(1);
}
#if _DEBUG
for (i=0; i<clique_size; i++)
printf("PMI_Get_clique_ranks[%d]=%d\n", i, clique_ranks[i]);
#endif
free(clique_ranks);
if ((rc = PMI_KVS_Get_name_length_max(&kvs_name_len)) != PMI_SUCCESS) {
printf("FAILURE: PMI_KVS_Get_name_length_max: %d, task %d\n",
rc, pmi_rank);
exit(1);
}
#if _DEBUG
printf("PMI_KVS_Get_name_length_max = %d\n", kvs_name_len);
#endif
kvs_name = malloc(kvs_name_len);
if ((rc = PMI_KVS_Get_my_name(kvs_name, kvs_name_len)) != PMI_SUCCESS) {
printf("FAILURE: PMI_KVS_Get_my_name: %d, task %d\n", rc,
pmi_rank);
exit(1);
}
#if _DEBUG
printf("PMI_KVS_Get_my_name = %s\n", kvs_name);
#endif
if ((rc = PMI_KVS_Get_key_length_max(&key_len)) != PMI_SUCCESS) {
printf("FAILURE: PMI_KVS_Get_key_length_max: %d, task %d\n",
rc, pmi_rank);
exit(1);
}
key = malloc(key_len);
if ((rc = PMI_KVS_Get_value_length_max(&val_len)) != PMI_SUCCESS) {
printf("FAILURE: PMI_KVS_Get_value_length_max: %d, task %d\n",
rc, pmi_rank);
exit(1);
}
#if _DEBUG
printf("PMI_KVS_Get_value_length_max = %d\n", val_len);
#endif
val = malloc(val_len);
/* Build and set some key=val pairs */
snprintf(key, key_len, "ATTR_1_%d", procid);
snprintf(val, val_len, "A%d", procid+OFFSET_1);
if ((rc = PMI_KVS_Put(kvs_name, key, val)) != PMI_SUCCESS) {
printf("FAILURE: PMI_KVS_Put(%s,%s,%s): %d, task %d\n",
kvs_name, key, val, rc, pmi_rank);
exit(1);
}
#if _DEBUG
printf("PMI_KVS_Put(%s,%s,%s)\n", kvs_name, key, val);
#endif
snprintf(key, key_len, "attr_2_%d", procid);
snprintf(val, val_len, "B%d", procid+OFFSET_2);
if ((rc = PMI_KVS_Put(kvs_name, key, val)) != PMI_SUCCESS) {
printf("FAILURE: PMI_KVS_Put(%s,%s,%s): %d, task %d\n",
kvs_name, key, val, rc, pmi_rank);
exit(1);
}
#if _DEBUG
printf("PMI_KVS_Put(%s,%s,%s)\n", kvs_name, key, val);
#endif
/* Sync KVS across all tasks */
if ((rc = PMI_KVS_Commit(kvs_name)) != PMI_SUCCESS) {
printf("FAILURE: PMI_KVS_Commit: %d, task %d\n", rc, pmi_rank);
exit(1);
}
#if _DEBUG
printf("PMI_KVS_Commit completed\n");
#endif
if ((rc = PMI_Barrier()) != PMI_SUCCESS) {
printf("FAILURE: PMI_Barrier: %d, task %d\n", rc, pmi_rank);
exit(1);
}
#if _DEBUG
printf("PMI_Barrier completed\n");
#endif
/* Now lets get all keypairs and validate */
for (i=0; i<pmi_size; i++) {
snprintf(key, key_len, "ATTR_1_%d", i);
if ((rc = PMI_KVS_Get(kvs_name, key, val, val_len))
!= PMI_SUCCESS) {
printf("FAILURE: PMI_KVS_Get(%s): %d, task %d\n",
key, rc, pmi_rank);
exit(1);
}
if ((val[0] != 'A')
|| ((atoi(&val[1])-OFFSET_1) != i)) {
printf("FAILURE: Bad keypair %s=%s, task %d\n",
key, val, pmi_rank);
exit(1);
}
#if _DEBUG
if ((pmi_size <= 8) && (pmi_rank == 0)) /* limit output */
printf("PMI_KVS_Get(%s,%s) %s\n", kvs_name, key, val);
#endif
snprintf(key, key_len, "attr_2_%d", i);
if ((rc = PMI_KVS_Get(kvs_name, key, val, val_len))
!= PMI_SUCCESS) {
printf("FAILURE: PMI_KVS_Get(%s): %d, task %d\n",
key, rc, pmi_rank);
exit(1);
}
if ((val[0] != 'B')
|| ((atoi(&val[1])-OFFSET_2) != i)) {
printf("FAILURE: Bad keypair %s=%s, task %d\n",
key,val, pmi_rank);
exit(1);
}
#if _DEBUG
if ((pmi_size <= 8) && (pmi_rank == 1)) /* limit output */
printf("PMI_KVS_Get(%s,%s) %s\n", kvs_name, key, val);
#endif
}
/* use iterator */
if ((rc = PMI_KVS_Iter_first(kvs_name, key, key_len, val,
val_len)) != PMI_SUCCESS) {
printf("FAILURE: PMI_KVS_iter_first: %d, task %d\n", rc,
pmi_rank);
exit(1);
}
for (i=0; ; i++) {
if (key[0] == '\0') {
if (i != (pmi_size * 2)) {
printf("FAILURE: PMI_KVS_iter_next "
"cycle count(%d, %d), task %d\n",
i, pmi_size, pmi_rank);
}
break;
}
#if _DEBUG
if ((pmi_size <= 8) && (pmi_rank == 1)) { /* limit output */
printf("PMI_KVS_Iter_next(%s,%d): %s=%s\n", kvs_name,
i, key, val);
}
#endif
if ((rc = PMI_KVS_Iter_next(kvs_name, key, key_len,
val, val_len)) != PMI_SUCCESS) {
printf("FAILURE: PMI_KVS_iter_next: %d, task %d\n",
rc, pmi_rank);
exit(1);
}
}
/* Build some more key=val pairs */
snprintf(key, key_len, "ATTR_3_%d", procid);
snprintf(val, val_len, "C%d", procid+OFFSET_1);
if ((rc = PMI_KVS_Put(kvs_name, key, val)) != PMI_SUCCESS) {
printf("FAILURE: PMI_KVS_Put(%s,%s,%s): %d, task %d\n",
kvs_name, key, val, rc, pmi_rank);
exit(1);
}
#if _DEBUG
printf("PMI_KVS_Put(%s,%s,%s)\n", kvs_name, key, val);
#endif
snprintf(key, key_len, "attr_4_%d", procid);
snprintf(val, val_len, "D%d", procid+OFFSET_2);
if ((rc = PMI_KVS_Put(kvs_name, key, val)) != PMI_SUCCESS) {
printf("FAILURE: PMI_KVS_Put(%s,%s,%s): %d, task %d\n",
kvs_name, key, val, rc, pmi_rank);
exit(1);
}
#if _DEBUG
printf("PMI_KVS_Put(%s,%s,%s)\n", kvs_name, key, val);
#endif
/* Sync KVS across all tasks */
if ((rc = PMI_KVS_Commit(kvs_name)) != PMI_SUCCESS) {
printf("FAILURE: PMI_KVS_Commit: %d, task %d\n", rc, pmi_rank);
exit(1);
}
#if _DEBUG
printf("PMI_KVS_Commit completed\n");
#endif
if ((rc = PMI_Barrier()) != PMI_SUCCESS) {
printf("FAILURE: PMI_Barrier: %d, task %d\n", rc, pmi_rank);
exit(1);
}
#if _DEBUG
printf("PMI_Barrier completed\n");
#endif
/* Now lets get some keypairs and validate */
for (i=0; i<pmi_size; i++) {
snprintf(key, key_len, "ATTR_1_%d", i);
if ((rc = PMI_KVS_Get(kvs_name, key, val, val_len))
!= PMI_SUCCESS) {
printf("FAILURE: PMI_KVS_Get(%s): %d, task %d\n",
key, rc, pmi_rank);
exit(1);
}
if ((val[0] != 'A')
|| ((atoi(&val[1])-OFFSET_1) != i)) {
printf("FAILURE: Bad keypair %s=%s, task %d\n",
key, val, pmi_rank);
exit(1);
}
#if _DEBUG
if ((pmi_size <= 8) && (pmi_rank == 1)) /* limit output */
printf("PMI_KVS_Get(%s,%s) %s\n", kvs_name, key, val);
#endif
snprintf(key, key_len, "attr_4_%d", i);
if ((rc = PMI_KVS_Get(kvs_name, key, val, val_len))
!= PMI_SUCCESS) {
printf("FAILURE: PMI_KVS_Get(%s): %d, task %d\n",
key, rc, pmi_rank);
exit(1);
}
if ((val[0] != 'D')
|| ((atoi(&val[1])-OFFSET_2) != i)) {
printf("FAILURE: Bad keypair %s=%s, task %d\n",
key,val, pmi_rank);
exit(1);
}
#if _DEBUG
if ((pmi_size <= 8) && (pmi_rank == 1)) /* limit output */
printf("PMI_KVS_Get(%s,%s) %s\n", kvs_name, key, val);
#endif
}
/* Replicate the very heavy load that MVAPICH2 puts on PMI
* This load exceeds that of MPICH2 by a very wide margin */
#if _DEBUG
printf("Starting %d iterations each with %d PMI_KVS_Put and \n"
" one each PMI_KVS_Commit and KVS_Barrier\n",
BARRIER_CNT, PUTS_PER_BARRIER);
fflush(stdout);
#endif
for (i=0; i<BARRIER_CNT; i++) {
for (j=0; j<PUTS_PER_BARRIER; j++) {
snprintf(key, key_len, "ATTR_%d_%d_%d", i, j, procid);
snprintf(val, val_len, "C%d", procid+OFFSET_1);
if ((rc = PMI_KVS_Put(kvs_name, key, val)) !=
PMI_SUCCESS) {
printf("FAILURE: PMI_KVS_Put(%s,%s,%s): "
"%d, task %d\n",
kvs_name, key, val, rc, pmi_rank);
exit(1);
}
}
if ((rc= PMI_KVS_Commit(kvs_name)) != PMI_SUCCESS) {
printf("FAILURE: PMI_KVS_Commit: %d, task %d\n",
rc, pmi_rank);
exit(1);
}
if ((rc = PMI_Barrier()) != PMI_SUCCESS) {
printf("FAILURE: PMI_Barrier: %d, task %d\n",
rc, pmi_rank);
exit(1);
}
/* Don't bother with PMI_KVS_Get as those are all local
* and do not put a real load on srun or the network */
}
#if _DEBUG
printf("Interative PMI calls successful\n");
#endif
/* create new keyspace and test it */
if ((rc = PMI_KVS_Create(kvs_name, kvs_name_len)) != PMI_SUCCESS) {
printf("FAILURE: PMI_KVS_Create: %d, task %d\n", rc, pmi_rank);
exit(1);
}
#if _DEBUG
printf("PMI_KVS_Create %s\n", kvs_name);
#endif
if ((rc = PMI_KVS_Put(kvs_name, "KVS_KEY", "KVS_VAL")) != PMI_SUCCESS) {
printf("FAILURE: PMI_KVS_Put: %d, task %d\n", rc, pmi_rank);
exit(1);
}
#if _DEBUG
printf("PMI_KVS_Put(%s,KVS_KEY,KVS_VAL)\n", kvs_name);
#endif
if ((rc = PMI_KVS_Get(kvs_name, "KVS_KEY", val, val_len)) !=
PMI_SUCCESS) {
printf("FAILURE: PMI_KVS_Get(%s, KVS_KEY): %d, task %d\n",
kvs_name, rc, pmi_rank);
exit(1);
}
#if _DEBUG
printf("PMI_KVS_Get(%s,%s) %s\n", kvs_name, "KVS_KEY", val);
#endif
if ((rc = PMI_KVS_Destroy(kvs_name)) != PMI_SUCCESS) {
printf("FAILURE: PMI_KVS_Destroy(%s): %d, task %d\n",
kvs_name, rc, pmi_rank);
exit(1);
}
if ((rc = PMI_KVS_Get(kvs_name, "KVS_KEY", val, val_len)) !=
PMI_ERR_INVALID_KVS) {
printf("FAILURE: PMI_KVS_Get(%s, KVS_KEY): %d, task %d\n",
kvs_name, rc, pmi_rank);
exit(1);
}
if ((rc = PMI_Finalize()) != PMI_SUCCESS) {
printf("FAILURE: PMI_Finalize: %d, task %d\n", rc, pmi_rank);
exit(1);
}
if (_DEBUG || (pmi_rank < 4)) {
gettimeofday(&tv2, NULL);
delta_t = (tv2.tv_sec - tv1.tv_sec) * 1000000;
delta_t += tv2.tv_usec - tv1.tv_usec;
snprintf(tv_str, sizeof(tv_str), "usec=%ld", delta_t);
printf("PMI test ran successfully, for task %d, %s\n",
pmi_rank, tv_str);
}
if (pmi_rank == 0) {
printf("NOTE: All failures reported, ");
printf("but only first four successes reported\n");
}
exit(0);
}
int MPIDI_CH3U_Check_for_failed_procs(void)
{
int mpi_errno = MPI_SUCCESS;
int pmi_errno;
int len;
char *kvsname;
MPIR_Group *prev_failed_group, *new_failed_group;
MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_MPIDI_CH3U_CHECK_FOR_FAILED_PROCS);
MPIR_FUNC_VERBOSE_ENTER(MPID_STATE_MPIDI_CH3U_CHECK_FOR_FAILED_PROCS);
/* FIXME: Currently this only handles failed processes in
comm_world. We need to fix hydra to include the pgid along
with the rank, then we need to create the failed group from
something bigger than comm_world. */
mpi_errno = MPIDI_PG_GetConnKVSname(&kvsname);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
#ifdef USE_PMI2_API
{
int vallen = 0;
pmi_errno = PMI2_KVS_Get(kvsname, PMI2_ID_NULL, "PMI_dead_processes", MPIDI_failed_procs_string, PMI2_MAX_VALLEN, &vallen);
MPIR_ERR_CHKANDJUMP(pmi_errno, mpi_errno, MPI_ERR_OTHER, "**pmi_kvs_get");
}
#else
pmi_errno = PMI_KVS_Get_value_length_max(&len);
MPIR_ERR_CHKANDJUMP(pmi_errno, mpi_errno, MPI_ERR_OTHER, "**pmi_kvs_get_value_length_max");
pmi_errno = PMI_KVS_Get(kvsname, "PMI_dead_processes", MPIDI_failed_procs_string, len);
MPIR_ERR_CHKANDJUMP(pmi_errno, mpi_errno, MPI_ERR_OTHER, "**pmi_kvs_get");
#endif
if (*MPIDI_failed_procs_string == '\0') {
/* there are no failed processes */
MPIDI_Failed_procs_group = MPIR_Group_empty;
goto fn_exit;
}
MPL_DBG_MSG_S(MPIDI_CH3_DBG_OTHER, TYPICAL, "Received proc fail notification: %s", MPIDI_failed_procs_string);
/* save reference to previous group so we can identify new failures */
prev_failed_group = MPIDI_Failed_procs_group;
/* Parse the list of failed processes */
MPIDI_CH3U_Get_failed_group(-2, &MPIDI_Failed_procs_group);
/* get group of newly failed processes */
mpi_errno = MPIR_Group_difference_impl(MPIDI_Failed_procs_group, prev_failed_group, &new_failed_group);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
if (new_failed_group != MPIR_Group_empty) {
mpi_errno = MPIDI_CH3I_Comm_handle_failed_procs(new_failed_group);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
mpi_errno = terminate_failed_VCs(new_failed_group);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
mpi_errno = MPIR_Group_release(new_failed_group);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
}
/* free prev group */
if (prev_failed_group != MPIR_Group_empty) {
mpi_errno = MPIR_Group_release(prev_failed_group);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
}
fn_exit:
MPIR_FUNC_VERBOSE_EXIT(MPID_STATE_MPIDI_CH3U_CHECK_FOR_FAILED_PROCS);
return mpi_errno;
fn_oom: /* out-of-memory handler for utarray operations */
MPIR_ERR_SET1(mpi_errno, MPI_ERR_OTHER, "**nomem", "**nomem %s", "utarray");
fn_fail:
goto fn_exit;
}
static int s1_init(void)
{
PMI_BOOL initialized;
int spawned;
int rc, ret = OPAL_ERROR;
int i, rank, lrank, nrank;
char *pmix_id, tmp[64];
opal_value_t kv;
char *str;
uint32_t ui32;
opal_process_name_t ldr;
char **localranks=NULL;
if (PMI_SUCCESS != (rc = PMI_Initialized(&initialized))) {
OPAL_PMI_ERROR(rc, "PMI_Initialized");
return OPAL_ERROR;
}
if (PMI_TRUE != initialized && PMI_SUCCESS != (rc = PMI_Init(&spawned))) {
OPAL_PMI_ERROR(rc, "PMI_Init");
return OPAL_ERROR;
}
// setup hash table
opal_pmix_base_hash_init();
// Initialize space demands
rc = PMI_KVS_Get_value_length_max(&pmix_vallen_max);
if (PMI_SUCCESS != rc) {
OPAL_PMI_ERROR(rc, "PMI_KVS_Get_value_length_max");
goto err_exit;
}
pmix_vallen_threshold = pmix_vallen_max * 3;
pmix_vallen_threshold >>= 2;
rc = PMI_KVS_Get_name_length_max(&pmix_kvslen_max);
if (PMI_SUCCESS != rc) {
OPAL_PMI_ERROR(rc, "PMI_KVS_Get_name_length_max");
goto err_exit;
}
rc = PMI_KVS_Get_key_length_max(&pmix_keylen_max);
if (PMI_SUCCESS != rc) {
OPAL_PMI_ERROR(rc, "PMI_KVS_Get_key_length_max");
goto err_exit;
}
// Initialize job environment information
pmix_id = (char*)malloc(pmix_vallen_max);
if (pmix_id == NULL) {
ret = OPAL_ERR_OUT_OF_RESOURCE;
goto err_exit;
}
/* Get domain id */
if (PMI_SUCCESS != (rc = PMI_Get_kvs_domain_id(pmix_id, pmix_vallen_max))) {
free(pmix_id);
goto err_exit;
}
/* get our rank */
ret = PMI_Get_rank(&rank);
if( PMI_SUCCESS != ret ) {
OPAL_PMI_ERROR(ret, "PMI_Get_rank");
free(pmix_id);
goto err_exit;
}
/* Slurm PMI provides the job id as an integer followed
* by a '.', followed by essentially a stepid. The first integer
* defines an overall job number. The second integer is the number of
* individual jobs we have run within that allocation. */
s1_pname.jobid = strtoul(pmix_id, &str, 10);
s1_pname.jobid = (s1_pname.jobid << 16) & 0xffff0000;
if (NULL != str) {
ui32 = strtoul(str, NULL, 10);
s1_pname.jobid |= (ui32 & 0x0000ffff);
}
ldr.jobid = s1_pname.jobid;
s1_pname.vpid = rank;
/* store our name in the opal_proc_t so that
* debug messages will make sense - an upper
* layer will eventually overwrite it, but that
* won't do any harm */
opal_proc_set_name(&s1_pname);
opal_output_verbose(2, opal_pmix_base_framework.framework_output,
"%s pmix:s1: assigned tmp name",
OPAL_NAME_PRINT(s1_pname));
OBJ_CONSTRUCT(&kv, opal_value_t);
kv.key = strdup(OPAL_PMIX_JOBID);
kv.type = OPAL_UINT32;
kv.data.uint32 = s1_pname.jobid;
if (OPAL_SUCCESS != (ret = opal_pmix_base_store(&OPAL_PROC_MY_NAME, &kv))) {
OPAL_ERROR_LOG(ret);
OBJ_DESTRUCT(&kv);
goto err_exit;
}
OBJ_DESTRUCT(&kv);
/* save it */
OBJ_CONSTRUCT(&kv, opal_value_t);
kv.key = strdup(OPAL_PMIX_RANK);
kv.type = OPAL_UINT32;
kv.data.uint32 = rank;
if (OPAL_SUCCESS != (ret = opal_pmix_base_store(&OPAL_PROC_MY_NAME, &kv))) {
OPAL_ERROR_LOG(ret);
OBJ_DESTRUCT(&kv);
goto err_exit;
}
OBJ_DESTRUCT(&kv);
pmix_kvs_name = (char*)malloc(pmix_kvslen_max);
if (pmix_kvs_name == NULL) {
ret = OPAL_ERR_OUT_OF_RESOURCE;
goto err_exit;
}
rc = PMI_KVS_Get_my_name(pmix_kvs_name, pmix_kvslen_max);
if (PMI_SUCCESS != rc) {
OPAL_PMI_ERROR(rc, "PMI_KVS_Get_my_name");
goto err_exit;
}
/* get our local proc info to find our local rank */
if (PMI_SUCCESS != (rc = PMI_Get_clique_size(&nlranks))) {
OPAL_PMI_ERROR(rc, "PMI_Get_clique_size");
return rc;
}
/* save the local size */
OBJ_CONSTRUCT(&kv, opal_value_t);
kv.key = strdup(OPAL_PMIX_LOCAL_SIZE);
kv.type = OPAL_UINT32;
kv.data.uint32 = nlranks;
if (OPAL_SUCCESS != (ret = opal_pmix_base_store(&OPAL_PROC_MY_NAME, &kv))) {
OPAL_ERROR_LOG(ret);
OBJ_DESTRUCT(&kv);
goto err_exit;
}
OBJ_DESTRUCT(&kv);
lrank = 0;
nrank = 0;
ldr.vpid = rank;
if (0 < nlranks) {
/* now get the specific ranks */
lranks = (int*)calloc(nlranks, sizeof(int));
if (NULL == lranks) {
rc = OPAL_ERR_OUT_OF_RESOURCE;
OPAL_ERROR_LOG(rc);
return rc;
}
if (PMI_SUCCESS != (rc = PMI_Get_clique_ranks(lranks, nlranks))) {
OPAL_PMI_ERROR(rc, "PMI_Get_clique_ranks");
free(lranks);
return rc;
}
/* note the local ldr */
ldr.vpid = lranks[0];
/* save this */
memset(tmp, 0, 64);
for (i=0; i < nlranks; i++) {
(void)snprintf(tmp, 64, "%d", lranks[i]);
opal_argv_append_nosize(&localranks, tmp);
if (rank == lranks[i]) {
lrank = i;
nrank = i;
}
}
str = opal_argv_join(localranks, ',');
opal_argv_free(localranks);
OBJ_CONSTRUCT(&kv, opal_value_t);
kv.key = strdup(OPAL_PMIX_LOCAL_PEERS);
kv.type = OPAL_STRING;
kv.data.string = str;
if (OPAL_SUCCESS != (ret = opal_pmix_base_store(&OPAL_PROC_MY_NAME, &kv))) {
OPAL_ERROR_LOG(ret);
OBJ_DESTRUCT(&kv);
goto err_exit;
}
OBJ_DESTRUCT(&kv);
}
/* save the local leader */
OBJ_CONSTRUCT(&kv, opal_value_t);
kv.key = strdup(OPAL_PMIX_LOCALLDR);
kv.type = OPAL_UINT64;
kv.data.uint64 = *(uint64_t*)&ldr;
if (OPAL_SUCCESS != (ret = opal_pmix_base_store(&OPAL_PROC_MY_NAME, &kv))) {
OPAL_ERROR_LOG(ret);
OBJ_DESTRUCT(&kv);
goto err_exit;
}
OBJ_DESTRUCT(&kv);
/* save our local rank */
OBJ_CONSTRUCT(&kv, opal_value_t);
kv.key = strdup(OPAL_PMIX_LOCAL_RANK);
kv.type = OPAL_UINT16;
kv.data.uint16 = lrank;
if (OPAL_SUCCESS != (ret = opal_pmix_base_store(&OPAL_PROC_MY_NAME, &kv))) {
OPAL_ERROR_LOG(ret);
OBJ_DESTRUCT(&kv);
goto err_exit;
}
OBJ_DESTRUCT(&kv);
/* and our node rank */
OBJ_CONSTRUCT(&kv, opal_value_t);
kv.key = strdup(OPAL_PMIX_NODE_RANK);
kv.type = OPAL_UINT16;
kv.data.uint16 = nrank;
if (OPAL_SUCCESS != (ret = opal_pmix_base_store(&OPAL_PROC_MY_NAME, &kv))) {
OPAL_ERROR_LOG(ret);
OBJ_DESTRUCT(&kv);
goto err_exit;
}
OBJ_DESTRUCT(&kv);
/* get universe size */
ret = PMI_Get_universe_size(&i);
if (PMI_SUCCESS != ret) {
OPAL_PMI_ERROR(ret, "PMI_Get_universe_size");
goto err_exit;
}
/* push this into the dstore for subsequent fetches */
OBJ_CONSTRUCT(&kv, opal_value_t);
kv.key = strdup(OPAL_PMIX_UNIV_SIZE);
kv.type = OPAL_UINT32;
kv.data.uint32 = i;
if (OPAL_SUCCESS != (ret = opal_pmix_base_store(&OPAL_PROC_MY_NAME, &kv))) {
OPAL_ERROR_LOG(ret);
OBJ_DESTRUCT(&kv);
goto err_exit;
}
OBJ_DESTRUCT(&kv);
/* get job size */
ret = PMI_Get_size(&i);
if (PMI_SUCCESS != ret) {
OPAL_PMI_ERROR(ret, "PMI_Get_size");
goto err_exit;
}
OBJ_CONSTRUCT(&kv, opal_value_t);
kv.key = strdup(OPAL_PMIX_JOB_SIZE);
kv.type = OPAL_UINT32;
kv.data.uint32 = i;
if (OPAL_SUCCESS != (ret = opal_pmix_base_store(&OPAL_PROC_MY_NAME, &kv))) {
OPAL_ERROR_LOG(ret);
OBJ_DESTRUCT(&kv);
goto err_exit;
}
OBJ_DESTRUCT(&kv);
/* get appnum */
ret = PMI_Get_appnum(&i);
if (PMI_SUCCESS != ret) {
OPAL_PMI_ERROR(ret, "PMI_Get_appnum");
goto err_exit;
}
OBJ_CONSTRUCT(&kv, opal_value_t);
kv.key = strdup(OPAL_PMIX_APPNUM);
kv.type = OPAL_UINT32;
kv.data.uint32 = i;
if (OPAL_SUCCESS != (ret = opal_pmix_base_store(&OPAL_PROC_MY_NAME, &kv))) {
OPAL_ERROR_LOG(ret);
OBJ_DESTRUCT(&kv);
goto err_exit;
}
OBJ_DESTRUCT(&kv);
/* increment the init count */
++pmix_init_count;
return OPAL_SUCCESS;
err_exit:
PMI_Finalize();
return ret;
}
static int mca_initialize_pmi_v1(void)
{
PMI_BOOL initialized;
int spawned;
int rc, ret = OPAL_ERROR;
if (PMI_SUCCESS != (rc = PMI_Initialized(&initialized))) {
OPAL_PMI_ERROR(rc, "PMI_Initialized");
return OPAL_ERROR;
}
if( PMI_TRUE != initialized && PMI_SUCCESS != (rc = PMI_Init(&spawned)) ) {
OPAL_PMI_ERROR(rc, "PMI_Init");
return OPAL_ERROR;
}
// Initialize space demands
rc = PMI_KVS_Get_value_length_max(&pmi_vallen_max);
if( PMI_SUCCESS != rc ) {
OPAL_PMI_ERROR(rc, "PMI_KVS_Get_value_length_max");
goto err_exit;
}
rc = PMI_KVS_Get_name_length_max(&pmi_kvslen_max);
if (PMI_SUCCESS != rc ) {
OPAL_PMI_ERROR(rc, "PMI_KVS_Get_name_length_max");
goto err_exit;
}
rc = PMI_KVS_Get_key_length_max(&pmi_keylen_max);
if( PMI_SUCCESS != rc ) {
OPAL_PMI_ERROR(rc, "PMI_KVS_Get_key_length_max");
goto err_exit;
}
// Initialize job environment information
rc = PMI_Get_rank(&pmi_rank);
if( PMI_SUCCESS != rc ) {
OPAL_PMI_ERROR(rc, "PMI_Get_rank");
return OPAL_ERROR;
}
rc = PMI_Get_universe_size(&pmi_usize);
if( PMI_SUCCESS != rc ) {
OPAL_PMI_ERROR(rc, "PMI_Get_universe_size");
goto err_exit;
}
rc = PMI_Get_size(&pmi_size);
if( PMI_SUCCESS != rc ) {
OPAL_PMI_ERROR(rc, "PMI_Get_size");
goto err_exit;
}
rc = PMI_Get_appnum(&pmi_appnum);
if( PMI_SUCCESS != rc ) {
OPAL_PMI_ERROR(rc, "PMI_Get_appnum");
goto err_exit;
}
pmi_kvs_name = (char*)malloc(pmi_kvslen_max);
if( pmi_kvs_name == NULL ){
ret = OPAL_ERR_OUT_OF_RESOURCE;
goto err_exit;
}
rc = PMI_KVS_Get_my_name(pmi_kvs_name,pmi_kvslen_max);
if( PMI_SUCCESS != rc ) {
OPAL_PMI_ERROR(rc, "PMI2_Job_GetId");
goto err_exit;
}
return OPAL_SUCCESS;
err_exit:
PMI_Finalize();
return ret;
}
